Information processing program, information processing method, and information processing device.

Abstract

This invention provides an information processing program, information processing method, and information processing device that improve the accuracy of protein energy inference. [Solution] A method for prediction using machine learning potential such as HDNNP (High-Dimensional Neural Network Potentials) 10, wherein protein structure information (ALA1, ALA2) is used as input data 20a, and training data 20 is obtained in which the difference from a reference energy identified based on the energy corresponding to the protein is the correct energy 20b, and a model that infers the energy of the protein is trained based on the training data.

Description

Technical Field

[0001] The present invention relates to an information processing program and the like.

Background Art

[0002] In developing the drug discovery process, it is important to analyze the energy of proteins. For example, methods for calculating the energy of proteins include a method using first-principles calculations, a method using classical force fields, and a method of predicting using machine learning potentials such as HDNNP (High-Dimensional Neural Network Potentials).

[0003] The method using first-principles calculations is characterized by high accuracy but high computational cost. The method using classical force fields is characterized by low computational cost but low accuracy. The method of predicting using machine learning potentials can predict the energy of proteins with a higher degree of freedom than classical force fields.

[0004] Here, a conventional technique for predicting the energy of a protein from the structural information of the protein using the HDNNP method will be described. In the conventional technique, HDNNP is trained using training data with the structural information of the protein as input data and the correct energy of the protein as correct data. [[ID=二十一]] [[ID=二十二]]

[0005] [[ID=二十三]] [[ID=二十四]]In the conventional technique, the energy of the protein to be evaluated is inferred by inputting the structural information of the protein to be evaluated into the trained HDNNP. [[ID=二十五]] [[ID=二十六]]

Prior Art Documents

Patent Documents

[0006] [[ID=三十三]] [[ID=三十四]] [[ID=三十五]]

Patent Document 1

Patent Document 2

Patent Document 3

[0007] However, the conventional techniques described above have the problem of being unable to improve the accuracy of protein energy inference.

[0008] In one aspect, the present invention aims to provide an information processing program, an information processing method, and an information processing device that can improve the accuracy of protein energy inference. [Means for solving the problem]

[0009] In the first approach, the computer performs the following steps: The computer takes the protein structure information as input data and obtains training data, which consists of the difference from a reference energy identified based on the energy corresponding to the protein as ground truth data. The computer then trains a model to infer the energy of the protein based on this training data. [Effects of the Invention]

[0010] This can improve the accuracy of inferring protein energy. [Brief explanation of the drawing]

[0011] [Figure 1] Figure 1 is a diagram illustrating the structure of a protein. [Figure 2] Figure 2 shows the energy characteristics of proteins. [Figure 3] Figure 3 shows an example of the configuration of HDNNP. [Figure 4]Figure 4 is a diagram illustrating the process of training HDNNP. [Figure 5] Figure 5 is Figure (1), which shows the inference results for the energy of the evaluated protein. [Figure 6] Figure 6 is Figure (2), which shows the inference results for the energy of the evaluated protein. [Figure 7] Figure 7 is a diagram illustrating the minimum energy and the difference. [Figure 8] Figure 8 shows the configuration of the HDNNP according to this embodiment. [Figure 9] Figure 9 shows an example of a training dataset related to this embodiment. [Figure 10] Figure 10 is a diagram illustrating the process of training the HDNNP according to this embodiment. [Figure 11] Figure 11 is a diagram (1) showing the inference accuracy of the present invention compared to the prior art. [Figure 12] Figure 12 is a figure (2) showing the inference accuracy of the present invention compared to the prior art. [Figure 13] Figure 13 shows the differences in the characteristics of the ground truth data between the prior art and the present invention. [Figure 14] Figure 14 is a functional block diagram showing the configuration of the information processing device according to this embodiment. [Figure 15] Figure 15 is a flowchart showing the processing procedure of the information processing device according to this embodiment. [Figure 16] Figure 16 shows an example of a computer hardware configuration that achieves similar functions to the information processing device in the embodiment. [Modes for carrying out the invention]

[0012] The following describes in detail, with reference to the drawings, embodiments of the information processing program, information processing method, and information processing apparatus disclosed in this application. However, this invention is not limited to these embodiments. [Examples]

[0013] Before describing the information processing device according to this embodiment, we will provide a more detailed explanation of the protein, the structure of HDNNP, HDNNP training, and the problems of the prior art.

[0014] First, let's explain proteins. Figure 1 is a diagram illustrating the structure of a protein. Proteins are composed of multiple amino acid residues. For example, protein 5 shown in Figure 1 is composed of approximately 3000 amino acid residues. One example of the constituent amino acid residues is shown in callout 6. Callout 6 contains three ALAs and two TYRs. Protein 5 also contains one each of GLU, PRO, PHE, and ARG.

[0015] This section explains the characteristics of protein energy. Figure 2 shows the characteristics of protein energy. In graph G1 shown in Figure 2, the vertical axis corresponds to the protein's energy, and the horizontal axis corresponds to the frame (time). As shown in graph G1, the protein's structure changes over time, and its energy also changes. The more stable the positions of the particles contained within the protein, the lower the energy becomes.

[0016] Next, we will explain the structure of HDNNP. HDNNP has a Neural Network (NN) set up for each residue type. Figure 3 shows an example of the structure of HDNNP. HDNNP10 shown in Figure 3 includes NN11 for ALA and NN12 for PRO.

[0017] NN11 for ALA includes an input layer 11a, a hidden layer 11b, and an output layer 11c. NN12 for PRO includes an input layer 12a, a hidden layer 12b, and an output layer 12c. The values ​​output from output layers 11c and 12c are output to the summation node 13.

[0018] Next, a conventional technique for training the HDNNP10 described in FIG. 3 will be described. FIG. 4 is a diagram for explaining the conventional technique for training the HDNNP. The apparatus for training the HDNNP is simply referred to as the "apparatus". The apparatus trains the HDNNP10 using the learning data 20. The learning data 20 includes input data 20a and correct answer data 20b.

[0019] The input data 20a includes the structure information of the protein 5a. For example, assume that the protein 5a includes ALA1, ALA2, PRO1, PRO2, PRO3, and PRO4 as amino acid residues. The correct answer data 20b is set with the correct energy "-5800" of the protein 5a. Note that the protein 5a indicates a state at a certain point in time in the protein 5.

[0020] The apparatus inputs the structure information (ALA1, ALA2) of ALA1 and ALA2 to the input layer 11a of the NN11 for ALA. (ALA1, ALA2) is an array. Due to the constraints of the description in the specification, "[" and "]" are replaced with "(" and ")" for description (the same applies to other arrays). The shape of (ALA1, ALA2) is (2, 64). As a result, from the output layer 11c of the NN11 for ALA, the energy of ALA (E ALA1 , E ALA2 ) is output. (E ALA1 , E ALA2 ) is an array and its shape is (2, 1).

[0021] The apparatus inputs the structure information (PRO1, PRO2, PRO3, PRO4) of PRO1 to PRO4 to the input layer 12a of the NN12 for PRO. (PRO1, PRO2, PRO3, PRO4) is an array and its shape is (4, 64). As a result, from the output layer 12c of the NN12 for PRO, the energy of PRO (E PRO1 , E PRO2 , E PRO3 , E PRO4 ) is output. (E PRO1 , E PRO2 , E PRO3 , E PRO4 ) is an array and its shape is (4, 1).

[0022] The combined node 13 is (E ALA1 ,E ALA2 ) and (E PRO1 ,E PRO2 ,E PRO3 ,E PRO4 The total energy E of all residues, which is the sum of (and). all The device calculates the energy E. all Then, the parameters of NN11 for ALA and NN12 for PRO are updated to minimize the error with the correct energy of "-5800". For example, the instrument uses backpropagation when updating parameters.

[0023] The device trains HDNNP10 by repeatedly performing the above process using multiple training data registered in the training dataset.

[0024] The evaluation of the trained HDNNP10 model uses evaluation data. This evaluation data includes structural information of proteins not used in training, and their correct energy values. In the following explanation, proteins not used in training are referred to as "evaluation proteins."

[0025] The device inputs the structural information of the protein to be evaluated into a trained HDNNP10. The closer the energy output from the HDNNP10 is to the correct energy of the evaluation data, the higher the inference accuracy of the trained HDNNP10.

[0026] Furthermore, the structural information of the protein input to HDNNP10 uses descriptors that numerically represent the characteristics of the protein's particle arrangement. In HDNNP, the particle arrangement around each particle is represented by descriptors using atomic-central symmetry functions (wACSFs).

[0027] In the descriptors for atomic centrospheric symmetry functions (wACSFs), G 2 i (Radial Symmetry Function) and G 4 i(Angular Symmetry Function) is used.

[0028] G 2 i G is defined by equation (1). 2 i The distance R between particle i and the other particle j is ij The contributions corresponding to are added. g(Z) included in equation (1) j ) is a function for weighting by particle type (in this embodiment, amino acid type), and is defined by equation (2). Z in equation (2) j f is the type of residue j, and Mj is the mass of residue j. C is the cutoff function, defined by equation (3). The cutoff function is used to calculate the symmetry function over radius R C This is a damping function for performing the operation within a certain range.

[0029]

number

[0030]

number

[0031]

number

[0032] G 4 i G is defined by equation (4). 4 i This is the angle θ between particle i and the two surrounding particles j and k. ijk Includes information on hyperparameters η and R. S By providing multiple symmetry functions with different values ​​for ζ and λ, we can address the information reduction of the symmetry function due to summation. j ,Z k) is a function used to weight the particles by type (in this case, type of amino acid), and is defined by equation (5). M in equation (5) j M k These are the masses of residues j and k.

[0033]

number

[0034]

number

[0035] Next, I will explain the challenges of conventional technology. For example, I will describe a case where HDNNP10 is trained based on structural information of approximately 3500 types of proteins, and the trained HDNNP10 is evaluated using approximately 180 types of evaluation proteins.

[0036] Figure 5 is Figure (1), which shows the inference results for the energy of the evaluated protein. The vertical axis of graph G2 in Figure 5 is the axis corresponding to the inferred value when the structural information of the evaluated protein is input into HDNNP10. The horizontal axis of graph G2 is the ground truth data (ground truth energy) of the evaluated protein.

[0037] One plot in Graph G2 shows the relationship between the inferred value of a certain evaluation protein based on a given structural information and the ground truth data. Multiple structural information sets exist for the same evaluation protein. For example, in Figure 5, inferences were performed for approximately 180 types of proteins, with 50 different structural information sets, resulting in approximately 9000 plots. The more plots that are close to the line L1, which shows x=y, the higher the inference accuracy of HDNNP10.

[0038] Figure 6 is Figure (2), which shows the inference results for the energy of the evaluated protein. The explanation of the vertical and horizontal axes in graph G3 shown in Figure 6 is the same as the explanation of the vertical and horizontal axes in graph G2 in Figure 5. In graph G3, the relationship between the inferred value and the ground truth data is plotted for each protein (pair of protein and structural information) after being normalized to -1 to 1.

[0039] For example, as explained in Figure 2, even the same protein can change in structure and energy over time. However, as shown in Figure 6, conventional techniques cannot accurately estimate the energy for each structural information of the protein being evaluated. In other words, they cannot accurately estimate the energy change due to time.

[0040] Next, the information processing device according to this embodiment will be described. In the following description, the information processing device according to this embodiment will be referred to as "information processing device 100". As described above, in the conventional technology, the correct energy of the protein was used directly as the correct data for the training data used during training. In contrast, the information processing device 100 uses the "minimum energy of the protein" and the "difference from the minimum energy" as the correct data for the training data used during training.

[0041] Figure 7 is a diagram illustrating the minimum energy and difference. Graph G4 in Figure 7 shows the energy change of a certain protein over time. The vertical axis of graph G4 corresponds to energy, and the horizontal axis corresponds to time (t). Line L2 shows the energy change of a certain protein over time.

[0042] In the example shown in Figure 7, the minimum energy of the protein is the energy E at time t1. t1 The difference is the minimum energy (energy E). t1 This is the area enclosed by line segment L3 passing through () and line L2.

[0043] Next, the configuration of the HDNNP used by the information processing device 100 will be described. Figure 8 is a diagram showing the configuration of the HDNNP according to this embodiment. As shown in Figure 8, the HDNNP according to this embodiment has NN set for each residue type. The HDNNP 50 shown in Figure 8 includes NN11 for ALA, NN12 for PRO, and combined nodes 51 and 52. In this embodiment, for convenience, only NN for ALA and PRO are used, but in reality, NN for other amino acids such as LYN and GLY may also be included.

[0044] The explanation for NN11 for ALA and NN12 for PRO is the same as the explanation for NN11 for ALA and NN12 for PRO described in Figure 3. Note that the output layer 11c of NN11 for ALA has a minimum energy E1 ALA The node that outputs the difference E2 ALA It may also include a node that outputs the minimum energy E1 PRO The node that outputs the difference E2 PRO It may include a node that outputs [something].

[0045] The output layer 11c of NN11 for ALA has a minimum energy E1 ALA This is output to the combined node 51. The output layer 12c of NN12 for PRO has a minimum energy E1 PRO Output this to the summation node 51.

[0046] The output layer 11c of NN11 for ALA is difference E2 ALA The output is sent to node 52. The output layer 12c of the NN12 for PRO is the difference E2 PRO Output this to the summation node 52.

[0047] Next, we will describe an example of the process by which the information processing device 100 trains the HDNNP50 described in Figure 8. First, we will describe an example of a training dataset used by the information processing device 100. Here, we will explain using the example of training with 3 samples for proteins A and B, which are composed of two ALAs and one PRO. Here, we will explain using the example of training with 3 samples for proteins A and B, but proteins A and B may contain amino acids other than ALA and PRO, and proteins A and B may be composed of amino acids other than ALA and PRO. Also, for example, protein A may be composed of ALA and PRO, and protein B may be composed of LYS and GLY, so proteins A and B may each contain different amino acids.

[0048] Figure 9 shows an example of a training dataset according to this embodiment. In the example shown in Figure 9, the training dataset 60 includes a descriptor 61 for protein A, ground truth data 62 for protein A, a descriptor 63 for protein B, and ground truth data 64 for protein B.

[0049] The protein A descriptor 61 contains the ALA descriptor 61a and the PRO descriptor 61b. The ALA descriptor 61a contains three descriptors for each of the two types of ALA. The PRO descriptor 61b contains three descriptors for one type of PRO.

[0050] The ground truth data for protein A includes three ground truth values ​​(E1 minimum energy, E2 difference).

[0051] When training HDNNP50 using the training dataset 60 in Figure 9, the input data and ground truth data pairs are as follows. For example, the ALA descriptor 61a "(23.4,45.2,54.2,...)" and (33.4,75.2,23.2,...)" and the PRO descriptor 61b "(61.4,23.2,54.2,...)" are input data. The ground truth data for protein A corresponding to this input data is "(-1000,20)".

[0052] The input data consists of ALA descriptor 61a "(74.4,42.2,4.2,...)" and (23.4,45.2,54.2,...)", and PRO descriptor 61b "(68.4,34.2,52.5,...)". The correct answer data for protein A corresponding to this input data is "(-1000,34)".

[0053] The input data consists of ALA descriptor 61a "(33.4,75.2,23.2,...)" and (74.4,42.2,4.2,...)" and PRO descriptor 61b "(36.4,26.2,34.7,...)". The correct data for protein A corresponding to this input data is "(-1000,0)".

[0054] While we will omit a detailed explanation of the ALA descriptor 63a and PRO descriptor 63b of protein B descriptor 63, the input data and the ground truth data are associated in the same way as the protein A descriptor 61.

[0055] Figure 10 is a diagram illustrating the process of training the HDNNP according to this embodiment. The information processing device 100 trains the HDNNP 50 using the training dataset 60. The information processing device 100 obtains the ALA descriptors 61a "(23.4,45.2,54.2,...), (33.4,75.2,23.2,...)" from the training dataset 60 and inputs them to the input layer 11a of the NN 11 for ALA, thereby obtaining the minimum energy E1 of the ALA. ALA And, difference E2 ALA The following is output. Minimum energy of ALA E1 ALA This is output to the aggregate node 51. ALA difference E2 ALA This is output to the aggregation node 52.

[0056] The information processing device 100 obtains the PRO descriptor 61b "(61.4,23.2,54.2,...)" from the training dataset 60 and inputs it to the input layer 12a of the NN12 for PRO, thereby determining the minimum energy E1 of PRO. PRO And, difference E2 PROThe following is output. PRO's minimum energy E1 PRO This is output to the combined node 51. PRO difference E2 PRO This is output to the aggregation node 52.

[0057] The combined node 51 is the minimum energy E1 of ALA. ALA And, the minimum energy of PRO E1 PRO E1 ALL The calculation is performed. The combined node 52 is the difference E2 of ALA. ALA And, PRO difference E2 PRO E2 ALL Calculate.

[0058] The information processing device 100 is E1 ALL The difference between the correct data "-1000" and E2 ALL The parameters for NN11 for ALA and NN12 for PRO are updated so that the difference between the correct answer value of "20" and the actual answer value becomes smaller.

[0059] The information processing device 100 performs the same processing as described above for pairs of other input data included in the descriptor 61 of protein A in the training dataset 60 and the ground truth data included in the ground truth data 62 for protein A, and updates the parameters of NN11 for ALA and NN12 for PRO.

[0060] Furthermore, the information processing device 100 performs the same processing as described above on pairs of other input data included in the descriptor 63 of protein B in the training dataset 60 and the ground truth data included in the ground truth data 64 of protein B, and updates the parameters of NN11 for ALA and NN12 for PRO.

[0061] The information processing device 100 repeatedly executes the above process until the termination condition is met. For example, the termination condition is that the number of epochs reaches a predetermined number. Alternatively, the termination condition is that the inference accuracy of HDNNP50 using the evaluation data is equal to or greater than the target accuracy.

[0062] For example, the evaluation data includes protein structure information (evaluation structure information), the minimum energy to be evaluated (evaluation minimum energy), and the difference (evaluation difference). The information processing device 100 inputs the evaluation structure information to the HDNNP 50 and estimates the minimum energy and the difference. The information processing device determines that the inference accuracy is equal to or better than the target accuracy if the difference between the estimated minimum energy and the evaluation minimum energy is less than the first threshold, and the difference between the estimated difference and the evaluation difference is less than the second threshold.

[0063] The process by which the information processing device 100 according to this embodiment trains the HDNNP 50 has been described above.

[0064] Next, we will explain the difference in inference accuracy when comparing the inference results of HDNNP50 trained using Figure 10 with the inference results of the prior art. Figure 11 is Figure (1) showing the inference accuracy of the present invention compared with the prior art. In Figure 11, we show the results of training using 80 types of proteins and evaluating the inference accuracy using the proteins used for training (so-called interpolation).

[0065] Let HDNNP trained using conventional technology be designated as HDNNP10, and HDNNP trained by the information processing device 100 be designated as HDNNP50. As described above, conventional technology uses the correct energy of the protein directly as the correct data. On the other hand, the information processing device 100 uses the "minimum energy of the protein" and the "difference from the minimum energy" as the correct data.

[0066] Graph G1-1 shows the relationship between the inferred values ​​obtained when the structural information of protein C is input into HDNNP10 and the correct data. Graph G1-2 shows the relationship between the inferred values ​​obtained when the structural information of protein C is input into HDNNP50 and the correct data.

[0067] Graph G2-1 shows the relationship between the inferred values ​​obtained when the structural information of protein D was input into HDNNP10 and the ground truth data. Graph G2-2 shows the relationship between the inferred values ​​obtained when the structural information of protein D was input into HDNNP50 and the ground truth data.

[0068] Graph G3-1 shows the relationship between the inferred values ​​obtained when the structural information of protein E was input into HDNNP10 and the correct data. Graph G3-2 shows the relationship between the inferred values ​​obtained when the structural information of protein E was input into HDNNP50 and the correct data.

[0069] For example, proteins C, D, and E are proteins included in the training dataset. Comparing graphs G1-1 and G1-2, graphs G2-1 and G2-2, and graphs G3-1 and G3-2, it can be seen that the inference accuracy of the present invention is better evaluated than that of the prior art.

[0070] Figure 12 is a figure (2) showing the inference accuracy of the present invention compared to the prior art. Figure 12 shows the results of training using 80 types of proteins and evaluating the inference accuracy using evaluation proteins that were not used in training (so-called extrapolation).

[0071] Graph G4-1 shows the relationship between the inferred value when the structural information of the evaluated protein X is input into HDNNP10 and the ground truth data. Graph G4-2 shows the relationship between the inferred value when the structural information of the evaluated protein X is input into HDNNP50 and the ground truth data.

[0072] Graph G5-1 shows the relationship between the inferred value when the structural information of the evaluated protein Y is input into HDNNP10 and the ground truth data. Graph G5-2 shows the relationship between the inferred value when the structural information of the evaluated protein Y is input into HDNNP50 and the ground truth data.

[0073] Graph G6-1 shows the relationship between the inferred value when the structural information of the evaluated protein Z is input into HDNNP10 and the ground truth data. Graph G6-2 shows the relationship between the inferred value when the structural information of the evaluated protein Z is input into HDNNP50 and the ground truth data.

[0074] For example, the evaluation proteins X, Y, and Z are proteins not included in the training dataset. Comparing graphs G4-1 and G4-2, G5-1 and G5-2, and G6-1 and G6-2, it can be seen that even when evaluating extrapolation, the inference accuracy of the present invention is better than that of the prior art.

[0075] Next, we will examine the differences between the characteristics of the ground truth data obtained using conventional technology and the characteristics of the ground truth data used in the present invention.

[0076] Figure 13 shows the differences in the characteristics of the ground truth data between the prior art and the present invention. Graph G10A in Figure 13 shows the change in energy over time in the ground truth data for each protein of the prior art. The vertical axis of G10A corresponds to energy, and the horizontal axis corresponds to time. For example, line L1A represents the ground truth data of protein A of the prior art. Line L1B represents the ground truth data of protein B of the prior art. Line L1C represents the ground truth data of protein C of the prior art. Note that, although the horizontal axis of G10A is set to time here, it may be an axis relating to other parameters.

[0077] As shown in graph G10A, the energy of each protein fluctuates on different scales. Therefore, when predicting the energy of the same type of protein, it is fine to use the relevant ground truth data for training. However, when predicting the energy of a different protein (evaluation protein), this can lead to a decrease in inference accuracy.

[0078] On the other hand, graph G10B in Figure 13 shows the change in energy over time in the ground truth data (here, the difference) for each protein of the present invention. The vertical axis of G10B corresponds to energy, and the horizontal axis corresponds to time. For example, line L2A is the ground truth data for protein A of the present invention. Line L2B is the ground truth data for protein B of the present invention. Line L2C is the ground truth data for protein C of the present invention.

[0079] As shown in graph G10B, the energy of each protein fluctuates on a similar scale. Therefore, even when predicting the energy of different types of proteins (evaluation proteins), the inference accuracy can be improved.

[0080] Next, an example of the configuration of the information processing device 100 described above will be explained. Figure 14 is a functional block diagram showing the configuration of the information processing device according to this embodiment. As shown in Figure 14, this information processing device 100 has a communication unit 110, an input unit 120, a display unit 130, a storage unit 140, and a control unit 150.

[0081] The communication unit 110 performs data communication with external devices, etc., via a network. The communication unit 110 may also receive training datasets 60, etc., from external devices.

[0082] The input unit 120 inputs various types of information to the control unit 150.

[0083] The display unit 130 displays the information output from the control unit 150.

[0084] The memory unit 140 contains HDNNP 50, a training dataset 60, and a sample DB 141. The memory unit 140 is a memory, etc.

[0085] HDNNP50 is a machine learning model that takes protein structure information as input and outputs the minimum energy and difference of the protein. Further details about HDNNP50 are the same as those explained in Figure 8, etc.

[0086] The training dataset 60 contains multiple training datasets for training HDNNP50. The input to the training dataset is protein structure information. The ground truth data for the training dataset consists of the minimum energy and difference data for the protein. Further details regarding the training dataset 60 are the same as those described in Figures 7 and 9.

[0087] Sample DB141 contains structural information for multiple sample proteins. The data structure of the protein structural information may be a descriptor.

[0088] The control unit 150 includes a generation unit 151, a training unit 152, and an inference unit 153. The control unit 150 is a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), etc.

[0089] The generation unit 151 generates a training dataset 60 based on the sample database 141. For example, the generation unit 151 obtains structural information of protein A from the sample database 141 and calculates the change in energy of protein A over time based on the structural information. For example, the generation unit 151 performs a Molecular Dynamics (MD) simulation to calculate the change in energy over a certain period of time.

[0090] The generation unit 151 identifies the minimum energy and the difference based on the calculated energy change over a certain period. The generation unit 151 registers the input data, which is the structural information of protein A, and the ground truth data corresponding to the minimum energy and the difference of protein A, into the training dataset 60.

[0091] The generation unit 151 generates a training dataset 60 by repeatedly performing the above process for other proteins registered in the sample DB 141.

[0092] Here, we have described the case where the generation unit 151 generates the training dataset 60 from the sample DB 141, but the training dataset 60 may be prepared in advance.

[0093] The training unit 152 uses the training dataset 60 to train the HDNNP 50 based on backpropagation. For example, the training unit 152 retrieves training data from the training dataset 60, inputs the input data contained in the training data into the HDNNP 50, and updates the parameters of the HDNNP 50 so that the output from the HDNNP 50 approaches the ground truth data. Further details regarding the training unit 152 are the same as those described in Figure 10.

[0094] The inference unit 153 uses the HDNNP 50, which has been trained by the training unit 152, to infer the energy of the protein. For example, the inference unit 153 inputs the structural information of the protein to be inferred into the HDNNP 50 and infers the minimum energy of the protein and the difference. The inference unit 153 infers the energy of the protein by summing the inferred minimum energy and the difference. The inference unit 153 outputs the inference result to the display unit 130 for display.

[0095] Next, an example of the processing procedure of the information processing device 100 according to this embodiment will be described. Figure 15 is a flowchart of the processing procedure of the information processing device according to this embodiment. As shown in Figure 15, the generation unit 151 of the information processing device 100 calculates the minimum energy and difference based on the structural information of the protein contained in the sample DB 141 and generates a training dataset 60 (step S101).

[0096] The training unit 152 of the information processing device 100 acquires training data from the training dataset 60 and trains HDNNP 50 (step S102). The training unit 152 evaluates HDNNP 50 based on the evaluation data (step S103). Note that the training dataset 60 in Figure 14 is divided into training data and evaluation data.

[0097] The training unit 154 determines whether the termination conditions are met (step S104). If the training unit 154 does not meet the termination conditions (step S104, No), it proceeds to step S102. On the other hand, if the training unit 154 does meet the termination conditions (step S104, Yes), it proceeds to step S105.

[0098] The inference unit 153 of the information processing device 100 acquires structural information of the protein to be inferred (step S105). The inference unit 153 inputs the structural information of the protein to be inferred into the trained HDNNP 50 and infers the minimum energy and the difference (step S106).

[0099] The inference unit 153 calculates the energy of the protein to be inferred by summing the minimum energy and the difference (step S107). The inference unit 153 outputs the calculation result (step S108).

[0100] Next, the effects of the information processing device 100 according to this embodiment will be described. The information processing device 100 takes protein structure information as input data and trains HDNNP50 based on training data in which the minimum energy and difference of the protein are ground truth data. As a result, it is possible to generate HDNNP50 with higher protein estimation accuracy than the conventional HDNNP10.

[0101] The information processing device 100 inputs the structural information of the protein to be inferred into the trained HDNNP 50, infers the minimum energy and difference of the protein to be inferred, and then infers the energy by summing the minimum energy and difference. This process can improve the inference accuracy, as explained in Figures 11 and 12.

[0102] By the way, the processing content of the information processing device 100 described above is just one example, and the information processing device 100 can perform other processing. For example, the information processing device 100 used "minimum energy of protein" and "difference from minimum energy" as ground truth data to be used as training data, but it is not limited to this. The information processing device 100 may also use "maximum energy of protein" and "difference from maximum energy," or "average energy of protein" and "difference from average energy" as ground truth data. The minimum energy, maximum energy, and average energy of protein correspond to the "reference energy." In the following explanation, the minimum energy, maximum energy, and average energy of protein will be referred to as the reference energy. Note that the reference energy is not limited to these; the median energy or mode energy may also be used.

[0103] Furthermore, while the information processing device 100 used a combination of "protein reference energy" and "difference from reference energy" as ground truth data for training, it is not limited to this. The information processing device 100 may also use only "difference from reference energy" as ground truth data for training.

[0104] Thus, when using only the "difference from the reference energy" as the ground truth data for training, the inference values ​​output from the HDNNP50 after training will also consist only of the difference from the reference energy.

[0105] Next, an example of a computer hardware configuration that achieves the same functions as the information processing device 100 described above will be explained. Figure 16 is a diagram showing an example of a computer hardware configuration that achieves the same functions as the information processing device in the embodiment.

[0106] As shown in Figure 16, the computer 200 includes a CPU 201 that performs various calculations, an input device 202 that receives data input from the user, and a display 203. The computer 200 also includes a communication device 204 and an interface device 205 that exchange data with external devices via a wired or wireless network. Furthermore, the computer 200 includes a RAM 206 for temporarily storing various information and a hard disk drive 207. Each of these devices 201 to 207 is connected to a bus 208.

[0107] The hard disk drive 207 contains a generation program 207a, a training program 207b, and an inference program 207c. The CPU 201 reads each of the programs 207a to 207c and loads them into the RAM 206.

[0108] The generation program 207a functions as the generation process 206a. The training program 207b functions as the training process 206b. The inference program 207c functions as the inference process 206c.

[0109] The processing of generation process 206a corresponds to the processing of generation unit 151. The processing of training process 206b corresponds to the processing of training unit 152. The processing of inference process 206c corresponds to the processing of inference unit 153.

[0110] Furthermore, programs 207a to 207c do not necessarily have to be stored on the hard disk drive 207 from the beginning. For example, each program could be stored on a "portable physical medium" such as a flexible disk (FD), CD-ROM, DVD, magneto-optical disk, or IC card inserted into the computer 200. Then, the computer 200 could read and execute each program 207a to 207c.

[0111] With regard to embodiments including each of the above examples, the following additional information is disclosed.

[0112] (Note 1) Using the structural information of the protein as input data, training data is obtained in which the difference between the reference energy identified based on the energy corresponding to the protein and the said energy is used as the ground truth data. Based on the aforementioned training data, a model is trained to infer the energy of the protein. An information processing program characterized by having a computer perform the processing.

[0113] (Note 2) The information processing program according to Note 1, wherein the correct data of the learning data further includes the reference energy, and the training process is characterized in that the model is trained based on the learning data in which the reference energy is further included in the correct data.

[0114] (Note 3) The information processing program according to Note 2, characterized in that the reference energy is one of the minimum energy, maximum energy, average energy, median energy, or mode energy identified based on the energy corresponding to the protein.

[0115] (Note 4) The information processing program according to Note 1, characterized in that it obtains structural information of the protein to be evaluated, and inputs the structural information of the protein to be evaluated into a model trained by the training process described above, thereby causing the computer to further perform a process to infer the energy of the protein to be evaluated.

[0116] (Note 5) The information processing program described in Note 1, characterized in that the training process involves training HDNNP (High-Dimensional Neural Network Potentials) as the model.

[0117] (Note 6) The information processing program described in any one of Notes 1 to 5, characterized in that the energy corresponding to the protein is the time-series energy of the protein.

[0118] (Note 7) Using the structural information of the protein as input data, training data is obtained in which the difference between the reference energy identified based on the energy corresponding to the protein and the said energy is used as the ground truth data. Based on the aforementioned training data, a model is trained to infer the energy of the protein. An information processing method characterized in that the processing is performed by a computer.

[0119] (Note 8) The information processing method according to Note 7, wherein the correct data of the learning data further includes the reference energy, and the training process is characterized in that the model is trained based on the learning data in which the reference energy is further included in the correct data.

[0120] (Note 9) The information processing method according to Note 8, characterized in that the reference energy is one of the minimum energy, maximum energy, average energy, median energy, or mode energy identified based on the energy corresponding to the protein.

[0121] (Note 10) The information processing method according to Note 7, characterized in that the computer further performs a process to infer the energy of the protein to be evaluated by obtaining structural information of the protein to be evaluated and inputting the structural information of the protein to be evaluated into a model trained by the training process described above.

[0122] (Note 11) The information processing method according to Note 7, characterized in that the training process involves training HDNNP (High-Dimensional Neural Network Potentials) as the model.

[0123] (Note 12) The information processing method according to any one of Notes 7 to 11, characterized in that the energy corresponding to the protein is the time-series energy of the protein.

[0124] (Note 13) Using the structural information of the protein as input data, and obtaining training data in which the difference between the reference energy identified based on the energy corresponding to the protein and the said energy is used as the ground truth data, Based on the aforementioned training data, a model is trained to infer the energy of the protein. An information processing device having a control unit that performs processing.

[0125] (Note 14) The information processing apparatus according to Note 13, wherein the correct data of the learning data further includes the reference energy, and the training process is characterized in that the model is trained based on the learning data in which the reference energy is further included in the correct data.

[0126] (Note 15) The information processing apparatus according to Note 14, characterized in that the reference energy is one of the minimum energy, maximum energy, or average energy identified based on the time series energy of the protein.

[0127] (Note 16) The information processing apparatus according to Note 13, characterized in that the control unit acquires structural information of the protein to be evaluated and inputs the structural information of the protein to be evaluated into the model trained by the training process, thereby further executing a process to infer the energy of the protein to be evaluated.

[0128] (Note 17) The information processing apparatus according to Note 13, characterized in that the training process involves training HDNNP (High-Dimensional Neural Network Potentials) as the model.

[0129] (Note 18) The information processing device according to any one of Notes 13 to 17, characterized in that the energy corresponding to the protein is the time-series energy of the protein. [Explanation of Symbols]

[0130] 100 Information Processing Devices 110 Communications Department 120 Input section 130 Display section 140 Storage section 150 Control Unit 151 Generation part 152 Training Department 153 Reasoning part

Claims

1. Using the structural information of a protein as input data, and obtaining training data where the difference between the reference energy identified based on the energy corresponding to the protein and the said energy is used as the ground truth data, Based on the aforementioned training data, a model is trained to infer the energy of the protein. An information processing program characterized by having a computer perform the processing.

2. The information processing program according to claim 1, wherein the correct data of the learning data further includes the reference energy, and the training process is characterized in that the model is trained based on the learning data in which the reference energy is further included in the correct data.

3. The information processing program according to claim 2, characterized in that the reference energy is one of the minimum energy, maximum energy, average energy, median energy, or mode energy identified based on the energy corresponding to the protein.

4. The information processing program according to claim 1, characterized in that it obtains structural information of the protein to be evaluated, and inputs the structural information of the protein to be evaluated into a model trained by the training process, thereby causing the computer to further perform a process to infer the energy of the protein to be evaluated.

5. The information processing program according to claim 1, characterized in that the training process involves training HDNNP (High-Dimensional Neural Network Potentials) as the model.

6. The information processing program according to any one of claims 1 to 5, characterized in that the energy corresponding to the protein is the time-series energy of the protein.

7. Using the structural information of a protein as input data, and obtaining training data where the difference between the reference energy identified based on the energy corresponding to the protein and the said energy is used as the ground truth data, Based on the aforementioned training data, a model is trained to infer the energy of the protein. An information processing method characterized in that the processing is performed by a computer.

8. Using the structural information of a protein as input data, and obtaining training data where the difference between the reference energy identified based on the energy corresponding to the protein and the said energy is used as the ground truth data, Based on the aforementioned training data, a model is trained to infer the energy of the protein. An information processing device having a control unit that performs processing.

Images